Model weight licence

A model-weight licence is the legal document governing what a downstream party may do with a released set of trained AI model parameters (“weights”) — separately from any licence covering the software used to train or run the model, and separately from any licensing constraints on the data the model was trained on. For a research institution, the model-weight licence is the document that actually determines whether downloading and using a given model (for research, for a service, or for further fine-tuning) is lawful, and on what terms — it is not settled by whether the weights are freely downloadable or the project calls itself “open.”

The three-license problem

A single released model can carry up to three legally distinct sets of terms, and an institution needs to check all three before assuming it may reuse a model freely:

• The weights licence — governs use, modification, and redistribution of the trained parameter file itself.
• The software licence — governs the training/inference code (often a standard permissive licence such as Apache-2.0 or MIT, even when the weights themselves carry restrictions).
• Training-data terms — separate constraints on the data used to train the model, which do not automatically pass through to the released weights and are frequently undisclosed entirely.

These three can diverge sharply on the same release: it is common for the inference code to be Apache-2.0 licensed while the weights themselves carry a bespoke, restriction-bearing licence.

Why “open source” is contested for model weights

The Open Source Initiative (OSI) published its Open Source AI Definition (OSAID) v1.0 in October 2024, extending the Open Source Definition’s four freedoms to AI systems. Under OSAID, a genuinely open-source AI system must give downstream users enough information about the training data, code, and architecture/parameters to substantially recreate the system — not just a freely downloadable weights file. By that standard, most widely used “open” models, including Meta’s Llama family, Google’s Gemma, and Alibaba’s Qwen, do not qualify as open source: their training data and training code are not disclosed, only the resulting weights. The industry term for this middle category is open-weight — weights are downloadable and usable under some licence, but the system is not reproducible from published components the way OSI-approved open-source software is. OSI has specifically criticised Meta’s framing of the Llama licences as “open source” as a form of “open-washing.”

The three common licence families

• Permissive open-source licences (Apache-2.0, MIT) — occasionally applied directly to weights (e.g. some Mistral AI releases), imposing no use-based restrictions at all. These are genuinely open by the OSI’s older Open Source Definition and, where training data/code are also disclosed, can meet the newer OSAID bar too.
• RAIL / OpenRAIL licences (Responsible AI Licenses, developed via the RAIL Initiative and used for releases including BigScience’s BLOOM) — permit free commercial and non-commercial reuse but attach behavioural use restrictions (e.g. prohibited-use lists covering things like generating disinformation or violating law) that apply to the model and to derivatives built from it. The RAIL Initiative’s own FAQ states that OpenRAIL licences are explicitly not open-source licences under the OSI definition, precisely because of these use-based restrictions.
• Bespoke community licences (the Llama Community License is the dominant example) — free for most commercial and research use, but with negotiated carve-outs. The Llama licences prohibit using Llama outputs to train competing large language models, and include a scale trigger: if a licensee’s products or services exceed 700 million monthly active users in the preceding calendar month, the free licence terminates and a separate commercial licence must be negotiated with Meta directly. Neither the Llama 2 nor Llama 3.x community licences are OSI-approved.

What this means for a research institution

Before an institution downloads and deploys a model’s weights — for a funded research project, an internal tool, or a service made available to students or the public — someone needs to actually read the specific licence attached to that specific model release, not assume “open” or “free to download” means unrestricted. Practical checkpoints:

• Check whether the licence imposes use-based restrictions (RAIL-style prohibited-use lists) that could conflict with the intended research application.
• Check for scale or field-of-use triggers (the Llama 700-million-MAU clause is the best-known example, but bespoke licences vary) before building anything intended to grow into production use.
• Check whether the licence prohibits using the model’s outputs to train a competing model — a common Llama-style restriction directly relevant to labs doing model-training research themselves.
• Do not assume the weights licence, the code licence, and the training-data terms are the same document, or that any one of them being permissive implies the others are too.

Where model provenance and disclosure also matter for an institution’s AI-governance obligations — for example under the EU AI Act’s transparency requirements — see the related entries on AI transparency markers, C2PA content provenance, the fundamental rights impact assessment, and the European AI Office, which are distinct from weight licensing but frequently arise in the same institutional AI-adoption review.

References

• Open Source Initiative, Open Source AI Definition (OSAID) v1.0, published October 2024 — opensource.org/ai.
• RAIL Initiative, “FAQ — Responsible AI Licenses (RAIL),” licenses.ai/faq-2, on OpenRAIL’s relationship to the OSI definition.
• Meta, Llama 3.1 and Llama 3.2 Community License Agreements, llama.com — monthly-active-user threshold and training-restriction clauses.
• Open Source Initiative, public commentary on Meta’s “open source” framing of Llama licences as open-washing.